Merge master.kernel.org:/pub/scm/linux/kernel/git/davej/agpgart
[linux-2.6] / drivers / infiniband / hw / ipath / ipath_srq.c
1 /*
2  * Copyright (c) 2006 QLogic, Inc. All rights reserved.
3  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33
34 #include <linux/err.h>
35 #include <linux/vmalloc.h>
36
37 #include "ipath_verbs.h"
38
39 /**
40  * ipath_post_srq_receive - post a receive on a shared receive queue
41  * @ibsrq: the SRQ to post the receive on
42  * @wr: the list of work requests to post
43  * @bad_wr: the first WR to cause a problem is put here
44  *
45  * This may be called from interrupt context.
46  */
47 int ipath_post_srq_receive(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
48                            struct ib_recv_wr **bad_wr)
49 {
50         struct ipath_srq *srq = to_isrq(ibsrq);
51         struct ipath_rwq *wq;
52         unsigned long flags;
53         int ret;
54
55         for (; wr; wr = wr->next) {
56                 struct ipath_rwqe *wqe;
57                 u32 next;
58                 int i;
59
60                 if ((unsigned) wr->num_sge > srq->rq.max_sge) {
61                         *bad_wr = wr;
62                         ret = -ENOMEM;
63                         goto bail;
64                 }
65
66                 spin_lock_irqsave(&srq->rq.lock, flags);
67                 wq = srq->rq.wq;
68                 next = wq->head + 1;
69                 if (next >= srq->rq.size)
70                         next = 0;
71                 if (next == wq->tail) {
72                         spin_unlock_irqrestore(&srq->rq.lock, flags);
73                         *bad_wr = wr;
74                         ret = -ENOMEM;
75                         goto bail;
76                 }
77
78                 wqe = get_rwqe_ptr(&srq->rq, wq->head);
79                 wqe->wr_id = wr->wr_id;
80                 wqe->num_sge = wr->num_sge;
81                 for (i = 0; i < wr->num_sge; i++)
82                         wqe->sg_list[i] = wr->sg_list[i];
83                 wq->head = next;
84                 spin_unlock_irqrestore(&srq->rq.lock, flags);
85         }
86         ret = 0;
87
88 bail:
89         return ret;
90 }
91
92 /**
93  * ipath_create_srq - create a shared receive queue
94  * @ibpd: the protection domain of the SRQ to create
95  * @attr: the attributes of the SRQ
96  * @udata: not used by the InfiniPath verbs driver
97  */
98 struct ib_srq *ipath_create_srq(struct ib_pd *ibpd,
99                                 struct ib_srq_init_attr *srq_init_attr,
100                                 struct ib_udata *udata)
101 {
102         struct ipath_ibdev *dev = to_idev(ibpd->device);
103         struct ipath_srq *srq;
104         u32 sz;
105         struct ib_srq *ret;
106
107         if (srq_init_attr->attr.max_wr == 0) {
108                 ret = ERR_PTR(-EINVAL);
109                 goto done;
110         }
111
112         if ((srq_init_attr->attr.max_sge > ib_ipath_max_srq_sges) ||
113             (srq_init_attr->attr.max_wr > ib_ipath_max_srq_wrs)) {
114                 ret = ERR_PTR(-EINVAL);
115                 goto done;
116         }
117
118         srq = kmalloc(sizeof(*srq), GFP_KERNEL);
119         if (!srq) {
120                 ret = ERR_PTR(-ENOMEM);
121                 goto done;
122         }
123
124         /*
125          * Need to use vmalloc() if we want to support large #s of entries.
126          */
127         srq->rq.size = srq_init_attr->attr.max_wr + 1;
128         srq->rq.max_sge = srq_init_attr->attr.max_sge;
129         sz = sizeof(struct ib_sge) * srq->rq.max_sge +
130                 sizeof(struct ipath_rwqe);
131         srq->rq.wq = vmalloc_user(sizeof(struct ipath_rwq) + srq->rq.size * sz);
132         if (!srq->rq.wq) {
133                 ret = ERR_PTR(-ENOMEM);
134                 goto bail_srq;
135         }
136
137         /*
138          * Return the address of the RWQ as the offset to mmap.
139          * See ipath_mmap() for details.
140          */
141         if (udata && udata->outlen >= sizeof(__u64)) {
142                 struct ipath_mmap_info *ip;
143                 __u64 offset = (__u64) srq->rq.wq;
144                 int err;
145
146                 err = ib_copy_to_udata(udata, &offset, sizeof(offset));
147                 if (err) {
148                         ret = ERR_PTR(err);
149                         goto bail_wq;
150                 }
151
152                 /* Allocate info for ipath_mmap(). */
153                 ip = kmalloc(sizeof(*ip), GFP_KERNEL);
154                 if (!ip) {
155                         ret = ERR_PTR(-ENOMEM);
156                         goto bail_wq;
157                 }
158                 srq->ip = ip;
159                 ip->context = ibpd->uobject->context;
160                 ip->obj = srq->rq.wq;
161                 kref_init(&ip->ref);
162                 ip->mmap_cnt = 0;
163                 ip->size = PAGE_ALIGN(sizeof(struct ipath_rwq) +
164                                       srq->rq.size * sz);
165                 spin_lock_irq(&dev->pending_lock);
166                 ip->next = dev->pending_mmaps;
167                 dev->pending_mmaps = ip;
168                 spin_unlock_irq(&dev->pending_lock);
169         } else
170                 srq->ip = NULL;
171
172         /*
173          * ib_create_srq() will initialize srq->ibsrq.
174          */
175         spin_lock_init(&srq->rq.lock);
176         srq->rq.wq->head = 0;
177         srq->rq.wq->tail = 0;
178         srq->limit = srq_init_attr->attr.srq_limit;
179
180         spin_lock(&dev->n_srqs_lock);
181         if (dev->n_srqs_allocated == ib_ipath_max_srqs) {
182                 spin_unlock(&dev->n_srqs_lock);
183                 ret = ERR_PTR(-ENOMEM);
184                 goto bail_wq;
185         }
186
187         dev->n_srqs_allocated++;
188         spin_unlock(&dev->n_srqs_lock);
189
190         ret = &srq->ibsrq;
191         goto done;
192
193 bail_wq:
194         vfree(srq->rq.wq);
195
196 bail_srq:
197         kfree(srq);
198
199 done:
200         return ret;
201 }
202
203 /**
204  * ipath_modify_srq - modify a shared receive queue
205  * @ibsrq: the SRQ to modify
206  * @attr: the new attributes of the SRQ
207  * @attr_mask: indicates which attributes to modify
208  * @udata: user data for ipathverbs.so
209  */
210 int ipath_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
211                      enum ib_srq_attr_mask attr_mask,
212                      struct ib_udata *udata)
213 {
214         struct ipath_srq *srq = to_isrq(ibsrq);
215         int ret = 0;
216
217         if (attr_mask & IB_SRQ_MAX_WR) {
218                 struct ipath_rwq *owq;
219                 struct ipath_rwq *wq;
220                 struct ipath_rwqe *p;
221                 u32 sz, size, n, head, tail;
222
223                 /* Check that the requested sizes are below the limits. */
224                 if ((attr->max_wr > ib_ipath_max_srq_wrs) ||
225                     ((attr_mask & IB_SRQ_LIMIT) ?
226                      attr->srq_limit : srq->limit) > attr->max_wr) {
227                         ret = -EINVAL;
228                         goto bail;
229                 }
230
231                 sz = sizeof(struct ipath_rwqe) +
232                         srq->rq.max_sge * sizeof(struct ib_sge);
233                 size = attr->max_wr + 1;
234                 wq = vmalloc_user(sizeof(struct ipath_rwq) + size * sz);
235                 if (!wq) {
236                         ret = -ENOMEM;
237                         goto bail;
238                 }
239
240                 /*
241                  * Return the address of the RWQ as the offset to mmap.
242                  * See ipath_mmap() for details.
243                  */
244                 if (udata && udata->inlen >= sizeof(__u64)) {
245                         __u64 offset_addr;
246                         __u64 offset = (__u64) wq;
247
248                         ret = ib_copy_from_udata(&offset_addr, udata,
249                                                  sizeof(offset_addr));
250                         if (ret) {
251                                 vfree(wq);
252                                 goto bail;
253                         }
254                         udata->outbuf = (void __user *) offset_addr;
255                         ret = ib_copy_to_udata(udata, &offset,
256                                                sizeof(offset));
257                         if (ret) {
258                                 vfree(wq);
259                                 goto bail;
260                         }
261                 }
262
263                 spin_lock_irq(&srq->rq.lock);
264                 /*
265                  * validate head pointer value and compute
266                  * the number of remaining WQEs.
267                  */
268                 owq = srq->rq.wq;
269                 head = owq->head;
270                 if (head >= srq->rq.size)
271                         head = 0;
272                 tail = owq->tail;
273                 if (tail >= srq->rq.size)
274                         tail = 0;
275                 n = head;
276                 if (n < tail)
277                         n += srq->rq.size - tail;
278                 else
279                         n -= tail;
280                 if (size <= n) {
281                         spin_unlock_irq(&srq->rq.lock);
282                         vfree(wq);
283                         ret = -EINVAL;
284                         goto bail;
285                 }
286                 n = 0;
287                 p = wq->wq;
288                 while (tail != head) {
289                         struct ipath_rwqe *wqe;
290                         int i;
291
292                         wqe = get_rwqe_ptr(&srq->rq, tail);
293                         p->wr_id = wqe->wr_id;
294                         p->num_sge = wqe->num_sge;
295                         for (i = 0; i < wqe->num_sge; i++)
296                                 p->sg_list[i] = wqe->sg_list[i];
297                         n++;
298                         p = (struct ipath_rwqe *)((char *) p + sz);
299                         if (++tail >= srq->rq.size)
300                                 tail = 0;
301                 }
302                 srq->rq.wq = wq;
303                 srq->rq.size = size;
304                 wq->head = n;
305                 wq->tail = 0;
306                 if (attr_mask & IB_SRQ_LIMIT)
307                         srq->limit = attr->srq_limit;
308                 spin_unlock_irq(&srq->rq.lock);
309
310                 vfree(owq);
311
312                 if (srq->ip) {
313                         struct ipath_mmap_info *ip = srq->ip;
314                         struct ipath_ibdev *dev = to_idev(srq->ibsrq.device);
315
316                         ip->obj = wq;
317                         ip->size = PAGE_ALIGN(sizeof(struct ipath_rwq) +
318                                               size * sz);
319                         spin_lock_irq(&dev->pending_lock);
320                         ip->next = dev->pending_mmaps;
321                         dev->pending_mmaps = ip;
322                         spin_unlock_irq(&dev->pending_lock);
323                 }
324         } else if (attr_mask & IB_SRQ_LIMIT) {
325                 spin_lock_irq(&srq->rq.lock);
326                 if (attr->srq_limit >= srq->rq.size)
327                         ret = -EINVAL;
328                 else
329                         srq->limit = attr->srq_limit;
330                 spin_unlock_irq(&srq->rq.lock);
331         }
332
333 bail:
334         return ret;
335 }
336
337 int ipath_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
338 {
339         struct ipath_srq *srq = to_isrq(ibsrq);
340
341         attr->max_wr = srq->rq.size - 1;
342         attr->max_sge = srq->rq.max_sge;
343         attr->srq_limit = srq->limit;
344         return 0;
345 }
346
347 /**
348  * ipath_destroy_srq - destroy a shared receive queue
349  * @ibsrq: the SRQ to destroy
350  */
351 int ipath_destroy_srq(struct ib_srq *ibsrq)
352 {
353         struct ipath_srq *srq = to_isrq(ibsrq);
354         struct ipath_ibdev *dev = to_idev(ibsrq->device);
355
356         spin_lock(&dev->n_srqs_lock);
357         dev->n_srqs_allocated--;
358         spin_unlock(&dev->n_srqs_lock);
359         if (srq->ip)
360                 kref_put(&srq->ip->ref, ipath_release_mmap_info);
361         else
362                 vfree(srq->rq.wq);
363         kfree(srq);
364
365         return 0;
366 }